Skip to contentSkip to conversation

package/libseccomp: circumvent uClibc-ng bug on x86_64

PATCH New unixmania at gmail.com 3 participants 6 messages Oct 19, 2019

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 package/libseccomp: circumvent uClibc-ng bug on x86_64

From: Carlos Santos <[email protected]>

On uClibc up to v1.0.31, syscall() for x86_64 is defined in libc/sysdeps/linux/x86_64/syscall.S as

syscall:
movq %rdi, %rax /* Syscall number -> rax. */
movq %rsi, %rdi /* shift arg1 - arg5. */
movq %rdx, %rsi
movq %rcx, %rdx
movq %r8, %r10
movq %r9, %r8
movq 8(%rsp),%r9 /* arg6 is on the stack. */
syscall /* Do the system call. */
cmpq $-4095, %rax /* Check %rax for error. */
jae __syscall_error /* Branch forward if it failed. */
ret /* Return to caller. */

And __syscall_error is defined in
libc/sysdeps/linux/x86_64/__syscall_error.c as

int __syscall_error(void) attribute_hidden;
int __syscall_error(void)

{
        register int err_no __asm__ ("%rcx");
        __asm__ ("mov %rax, %rcx\n\t"
                 "neg %rcx");
        __set_errno(err_no);
        return -1;
}

Notice that __syscall_error returns -1 as a 32-bit int in %rax, a 64-bit register i.e. 0x00000000ffffffff (decimal 4294967295). When this value is compared to -1 in _sys_chk_seccomp_flag_kernel() the result is false, leading the function to always return 0.

Prevent the error by coercing the return value of syscall() to int in a temporary variable before comparing it to -1. We could use just an (int) cast but the variable makes the code more readable and the machine code generated by the compiler is the same in both cases.

All other syscall() invocations were inspected and they either already coerce the result to int or do not compare it to -1.

The same problem probably occurs on other 64-bit systems but so far only x86_64 was tested.

A bug report is being submitted to uClibc.

Upstream status: https://github.com/seccomp/libseccomp/pull/175

---
 ...n-uClibc-ng-syscall-on-x86_64-system.patch | 80 +++++++++++++++++++
 1 file changed, 80 insertions(+)
 create mode 100644 package/libseccomp/0002-Circumvent-bug-in-uClibc-ng-syscall-on-x86_64-system.patch
package/libseccomp/0002-Circumvent-bug-in-uClibc-ng-syscall-on-x86_64-system.patch new file +80 -0
new file mode 100644
index 0000000000..a7151d840f
@@ -0,0 +1,80 @@
+From 613e601bb4b50dc359b41f162a5b629449e4bbea Mon Sep 17 00:00:00 2001
+From: Carlos Santos <[email protected]>
+Date: Fri, 18 Oct 2019 22:02:49 -0300
+Subject: [PATCH] Circumvent bug in uClibc-ng syscall() on x86_64 systems
+
+On uClibc up to v1.0.31, syscall() for x86_64 is defined in
+libc/sysdeps/linux/x86_64/syscall.S as
+
+syscall:
+        movq %rdi, %rax         /* Syscall number -> rax.  */
+        movq %rsi, %rdi         /* shift arg1 - arg5.  */
+        movq %rdx, %rsi
+        movq %rcx, %rdx
+        movq %r8, %r10
+        movq %r9, %r8
+        movq 8(%rsp),%r9        /* arg6 is on the stack.  */
+        syscall                 /* Do the system call.  */
+        cmpq $-4095, %rax       /* Check %rax for error.  */
+        jae __syscall_error     /* Branch forward if it failed.  */
+        ret                     /* Return to caller.  */
+
+And __syscall_error is defined in
+libc/sysdeps/linux/x86_64/__syscall_error.c as
+
+int __syscall_error(void) attribute_hidden;
+int __syscall_error(void)
+{
+	register int err_no __asm__ ("%rcx");
+	__asm__ ("mov %rax, %rcx\n\t"
+	         "neg %rcx");
+	__set_errno(err_no);
+	return -1;
+}
+
+Notice that __syscall_error returns -1 as a 32-bit int in %rax, a 64-bit
+register i.e. 0x00000000ffffffff (decimal 4294967295). When this value
+is compared to -1 in _sys_chk_seccomp_flag_kernel() the result is false,
+leading the function to always return 0.
+
+Prevent the error by coercing the return value of syscall() to int in a
+temporary variable before comparing it to -1. We could use just an (int)
+cast but the variable makes the code more readable and the machine code
+generated by the compiler is the same in both cases.
+
+All other syscall() invocations were inspected and they either already
+coerce the result to int or do not compare it to -1.
+
+The same problem probably occurs on other 64-bit systems but so far only
+x86_64 was tested.
+
+A bug report is being submitted to uClibc.
+
+Signed-off-by: Carlos Santos <[email protected]>
+---
+ src/system.c | 8 +++++---
+ 1 file changed, 5 insertions(+), 3 deletions(-)
+
+diff --git a/src/system.c b/src/system.c
+index 8e5aafc..811b401 100644
+--- a/src/system.c
++++ b/src/system.c
+@@ -215,10 +215,12 @@ static int _sys_chk_seccomp_flag_kernel(int flag)
+ 	/* this is an invalid seccomp(2) call because the last argument
+ 	 * is NULL, but depending on the errno value of EFAULT we can
+ 	 * guess if the filter flag is supported or not */
+-	if (sys_chk_seccomp_syscall() == 1 &&
+-	    syscall(_nr_seccomp, SECCOMP_SET_MODE_FILTER, flag, NULL) == -1 &&
+-	    errno == EFAULT)
++	int rc;
++	if (sys_chk_seccomp_syscall() == 1) {
++	    rc = syscall(_nr_seccomp, SECCOMP_SET_MODE_FILTER, flag, NULL);
++	    if (rc == -1 && errno == EFAULT)
+ 		return 1;
++	}
+ 
+ 	return 0;
+ }
+-- 
+2.18.1
+
Signature · 1 line
2.18.1
#1 lore raw

PATCH 1/1 package/libseccomp: circumvent uClibc-ng bug on x86_64

Hello,

Quoted 2 lines · On Fri, 18 Oct 2019 22:15:55 -0300 unixmania at gmail.com wrote:

On uClibc up to v1.0.31, syscall() for x86_64 is defined in libc/sysdeps/linux/x86_64/syscall.S as

If it is "up to v1.0.31", then the problem is fixed in uClibc-ng. Why would we want to carry a work-around in libsseccomp ? To support older uClibc-ng pre-built toolchains ?

Quoted 4 lines

The same problem probably occurs on other 64-bit systems but so far only x86_64 was tested.

A bug report is being submitted to uClibc.

But you said above it's already fixed in uClibc-ng, so ?

Thomas

Signature · 3 lines
Thomas Petazzoni, CTO, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com
#2 lore raw

PATCH 1/1 package/libseccomp: circumvent uClibc-ng bug on x86_64

Quoted 9 lines

Em 19 de out de 2019, ?(s) 09:17, Thomas Petazzoni <[email protected]> escreveu:

?Hello,

Quoted 3 lines

On Fri, 18 Oct 2019 22:15:55 -0300
unixmania at gmail.com wrote:

Quoted 6 lines
Quoted 2 lines

On uClibc up to v1.0.31, syscall() for x86_64 is defined in libc/sysdeps/linux/x86_64/syscall.S as

If it is "up to v1.0.31", then the problem is fixed in uClibc-ng. Why would we want to carry a work-around in libsseccomp ? To support older uClibc-ng pre-built toolchains ?

It?s not fixed in uClibc-ng, since there no version newer than 1.0.31.

Quoted 7 lines
Quoted 5 lines

The same problem probably occurs on other 64-bit systems but so far only x86_64 was tested.

A bug report is being submitted to uClibc.

But you said above it's already fixed in uClibc-ng, so ?

It?s not.

Quoted 6 lines

Thomas

Signature · 3 lines
Thomas Petazzoni, CTO, Bootlin
Embedded Linux and Kernel engineering
https://

Carlos Santos <[email protected]>

#3 lore raw

PATCH 1/1 package/libseccomp: circumvent uClibc-ng bug on x86_64

Quoted 5 lines · On Sat, 19 Oct 2019 10:01:52 -0300 Carlos Santos <[email protected]> wrote:
Quoted 3 lines

If it is "up to v1.0.31", then the problem is fixed in uClibc-ng. Why would we want to carry a work-around in libsseccomp ? To support older uClibc-ng pre-built toolchains ?

It?s not fixed in uClibc-ng, since there no version newer than 1.0.31.

There is a 1.0.32, which Buildroot uses since a few days, hence when you said "up to 1.0.31", I thought it was fixed.

Thomas

Signature · 3 lines
Thomas Petazzoni, CTO, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com
#4 lore raw

PATCH 1/1 package/libseccomp: circumvent uClibc-ng bug on x86_64

Quoted 4 lines · Thomas Petazzoni

On Sat, 19 Oct 2019 10:01:52 -0300
Carlos Santos <[email protected]> wrote:

Quoted 8 lines
Quoted 5 lines
Quoted 3 lines

If it is "up to v1.0.31", then the problem is fixed in uClibc-ng. Why would we want to carry a work-around in libsseccomp ? To support older uClibc-ng pre-built toolchains ?

It?s not fixed in uClibc-ng, since there no version newer than 1.0.31.

There is a 1.0.32, which Buildroot uses since a few days, hence when you said "up to 1.0.31", I thought it was fixed.

Ah, I was not aware of that. I just checked and it still contains the error, so the workaround in libseccomp is still necessary.

Signature · 1 line
Carlos Santos <[email protected]>
#5 lore raw

PATCH 1/1 package/libseccomp: circumvent uClibc-ng bug on x86_64

In reply to unixmania at gmail.com
Quoted 31 lines · unixmania at gmail.com

From: Carlos Santos <[email protected]>

On uClibc up to v1.0.31, syscall() for x86_64 is defined in libc/sysdeps/linux/x86_64/syscall.S as

syscall:
movq %rdi, %rax /* Syscall number -> rax. */
movq %rsi, %rdi /* shift arg1 - arg5. */
movq %rdx, %rsi
movq %rcx, %rdx
movq %r8, %r10
movq %r9, %r8
movq 8(%rsp),%r9 /* arg6 is on the stack. */
syscall /* Do the system call. */
cmpq $-4095, %rax /* Check %rax for error. */
jae __syscall_error /* Branch forward if it failed. */
ret /* Return to caller. */

And __syscall_error is defined in
libc/sysdeps/linux/x86_64/__syscall_error.c as

int __syscall_error(void) attribute_hidden;
int __syscall_error(void)

{
        register int err_no __asm__ ("%rcx");
        __asm__ ("mov %rax, %rcx\n\t"
                 "neg %rcx");
        __set_errno(err_no);
        return -1;
}
Quoted 8 lines

Notice that __syscall_error returns -1 as a 32-bit int in %rax, a 64-bit register i.e. 0x00000000ffffffff (decimal 4294967295). When this value is compared to -1 in _sys_chk_seccomp_flag_kernel() the result is false, leading the function to always return 0.

Prevent the error by coercing the return value of syscall() to int in a temporary variable before comparing it to -1. We could use just an (int) cast but the variable makes the code more readable

Actually, I disagree with that. The intermediate variable hides the cast, and
the cast is the essential thing here. So if I were upstream, I'd want the explicit cast, *and* a big fat comment above it to explain why it is needed.

For us, however, that reasoning doesn't really fly because we see the full
patch, not the source code.

Quoted 6 lines

and the machine code
generated by the compiler is the same in both cases.

All other syscall() invocations were inspected and they either already coerce the result to int or do not compare it to -1.

Quoted 8 lines

The same problem probably occurs on other 64-bit systems but so far only x86_64 was tested.

A bug report is being submitted to uClibc.

Upstream status: https://github.com/seccomp/libseccomp/pull/175

I updated both the commit message and the patch to refer to uClibc 1.0.32
instead and applied to master, thanks.

Regards,
Arnout